Committee considers ‘ingredients label’ for AI training data; supporters and tech groups urge clarity on trade secrets

The committee heard House Bill 2,503 on Jan. 21, which would require developers publishing generative AI systems to post high-level documentation about training datasets — sources, approximate data point counts, whether copyrighted material was included, and whether datasets were licensed or purchased — before making systems publicly available in Washington.

Sponsors described the requirement as an 'ingredients label' that helps researchers evaluate bias and helps consumers and businesses understand the risks and provenance of AI systems. The staff summary and sponsor remarks stressed the requirement is high-level, not granular, and includes carve-outs to protect trade secrets and to exempt some systems (for example certain FDA-regulated systems).

Tech industry groups signaled conditional support but pressed for clearer trade-secret protections and pointed to California’s recently enacted law as a comparison point; Chamber of Progress and TechNet said enforcement design (private rights of action versus administrative enforcement) and clarity about which models are covered are critical. Privacy and safety advocates recounted findings that some widely used datasets have included illicit or abusive material and said stronger disclosure supports accountability and enforcement under the Consumer Protection Act.

Witnesses encouraged careful amendment drafting to avoid chilling beneficial AI uses in health care or research while ensuring disclosures are meaningful and enforceable. The committee closed the hearing and requested follow-up and amendment suggestions.